Langchain experiment

[marcnnn]

Some experiments with Langchain.

With a pachted Langchain version where:

field :receiver, :string

is added in LangChain.Message in message.ex

PR just to create visibility and a place to discuss.

[nileshtrivedi]

Amazing! 👍

I shared some notes on supporting langchain here. This is definitely on the right track. 👍 I am just waiting for langchain library v0.3.0 (which fixes tool calling for Gemini models) to be released before switching over to it.

[nileshtrivedi]

I have started merging langchain 0.3.0 in this commit.

I'm also simplifying things a bit by keeping thread.chat_history in chronological order so that it does not need to be reversed before sending to LLMChain. Ollama is now the default model and all scripts seem to be working fine.

Since I'm using the non-patched version of langchain, message "receiver" is no longer tracked in the thread. Instead of using a patched version, I am hoping that langchain can allow us to associate some arbitrary metadata per-message to support agent frameworks.